brightertiger/garbled-text-detector

Garbled Text Detector

Detect garbled, corrupted, or malformed text with 96.2% accuracy. Perfect for data quality assessment, PDF parsing validation, and text preprocessing pipelines.

Quick Start

Installation

pip install transformers torch

Basic Usage

The simplest way to use the model:

from transformers import pipeline

classifier = pipeline("text-classification", model="brightertiger/garbled-text-detector")

result = classifier("Your text here")[0]
print(f"Label: {result['label']}, Confidence: {result['score']:.2%}")

Batch Processing

from transformers import pipeline

classifier = pipeline("text-classification", model="brightertiger/garbled-text-detector")

texts = [
    "This is normal, well-formed text.",
    "H3ll0 w0rld! Th1s 1s g4rbl3d t3xt.",
    "The quick brown fox jumps over the lazy dog."
]

results = classifier(texts)

for text, result in zip(texts, results):
    print(f"{text[:50]:50} -> {result['label']} ({result['score']:.2%})")

Manual Model Loading (Advanced)

If you need more control:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("brightertiger/garbled-text-detector")
model = AutoModelForSequenceClassification.from_pretrained("brightertiger/garbled-text-detector")

text = "Your text here"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=1)
    prediction = torch.argmax(probs, dim=1).item()
    confidence = torch.max(probs).item()

label = model.config.id2label[prediction]
print(f"Prediction: {label} (confidence: {confidence:.2%})")

What It Does

This model classifies text into two categories:

Normal (Label 0): Clean, well-formed, readable text

• Example: "The company reported revenue of $1.2 million in Q4."

Garbled (Label 1): Corrupted, malformed, or unreadable text

• Example: "The%âãÏÓcompany%âãÏÓreported$1.2%âãÏÓmillion"

Common Use Cases

• Data Quality: Filter corrupted text from datasets before training
• PDF Parsing: Validate text extraction quality from PDFs
• OCR Validation: Check if OCR output is readable
• Encoding Detection: Identify character encoding issues
• Content Filtering: Remove malformed user-generated content

What It Handles

✅ Normal Text: Including technical content with:

• Mathematical symbols: ∑, ∫, ±, ≤, α, β
• Currency symbols: $, €, £, ¥
• Legal citations and references
• Formatted tables and structured data

✅ Detects Corruption:

• Binary data interpretation
• Encoding corruption
• Mixed character sets
• Non-ASCII noise

Performance

Metric	Score
Accuracy	96.2%
Precision	92.5%
Recall	95.4%
F1-Score	93.9%
ROC-AUC	99.1%

Inference Performance (CPU):

• Throughput: ~70 samples/second
• Latency: ~14ms per sample
• Model Size: 55MB
• Device: CPU optimized with dynamic quantization

Confusion Matrix (6,646 validation samples):

• True Negatives: 4,417 (Normal → Normal)
• False Positives: 171 (Normal → Garbled)
• False Negatives: 86 (Garbled → Normal)
• True Positives: 1,972 (Garbled → Garbled)

Model Details

Architecture:

• Base Model: TinyBERT (huawei-noah/TinyBERT_General_4L_312D)
• Model Type: AutoModelForSequenceClassification
• Parameters: 14.4M
• Max Sequence Length: 512 tokens
• Task: Binary classification (Normal vs Garbled)

Model Specifications:

• Framework: PyTorch + Transformers
• Model Size: 55MB
• Input: Text (up to 512 tokens)
• Output: Binary classification with confidence scores
• Labels: NORMAL (0) and GARBLED (1)

Training Data:

• Synthetic data generated with LLMs simulating real-world PDF parsing scenarios
• Includes diverse corruption patterns and edge cases
• Balanced dataset with both normal and garbled examples

Example Outputs

from transformers import pipeline

classifier = pipeline("text-classification", model="brightertiger/garbled-text-detector")

# Normal text
classifier("The quarterly earnings report shows strong growth.")
# → [{'label': 'NORMAL', 'score': 0.998}]

# Technical content (correctly identified as normal)
classifier("Calculate ROI = (Gain - Cost) / Cost × 100%")
# → [{'label': 'NORMAL', 'score': 0.973}]

# Currency symbols (correctly identified as normal)
classifier("Exchange rates: USD$1.00, EUR€0.85, GBP£0.73")
# → [{'label': 'NORMAL', 'score': 0.961}]

# Garbled text
classifier("PDF-1.4%âãÏÓ1 0 obj<< /Type /Catalog")
# → [{'label': 'GARBLED', 'score': 0.992}]

# Encoding corruption
classifier("Revenue%âãÏÓ$1,250,000%âãÏÓProfit%âãÏÓ$890,000")
# → [{'label': 'GARBLED', 'score': 0.987}]

Integration Examples

Data Cleaning Pipeline

from transformers import pipeline

def clean_dataset(texts):
    classifier = pipeline("text-classification", model="brightertiger/garbled-text-detector")
    results = classifier(texts)
    
    clean_texts = [
        text for text, result in zip(texts, results)
        if result['label'] == "NORMAL" and result['score'] > 0.9
    ]
    
    return clean_texts

Quality Control

from transformers import pipeline

def validate_text_quality(text, threshold=0.85):
    classifier = pipeline("text-classification", model="brightertiger/garbled-text-detector")
    result = classifier(text)[0]
    
    if result['label'] == "GARBLED" and result['score'] > threshold:
        return {
            "status": "rejected",
            "reason": "garbled_text",
            "confidence": result['score']
        }
    
    return {"status": "approved", "confidence": result['score']}

PDF Extraction Validation

from transformers import pipeline

def validate_pdf_extraction(pdf_text):
    classifier = pipeline("text-classification", model="brightertiger/garbled-text-detector")
    chunks = [pdf_text[i:i+1000] for i in range(0, len(pdf_text), 1000)]
    
    results = classifier(chunks)
    garbled_count = sum(1 for r in results if r['label'] == "GARBLED")
    garbled_ratio = garbled_count / len(chunks)
    
    if garbled_ratio > 0.2:
        return {"quality": "poor", "garbled_ratio": garbled_ratio}
    
    return {"quality": "good", "garbled_ratio": garbled_ratio}

API Reference

Using with Transformers Pipeline

from transformers import pipeline

# Load the model
classifier = pipeline(
    "text-classification", 
    model="brightertiger/garbled-text-detector"
)

# Single prediction
result = classifier("Your text here")[0]
print(f"{result['label']}: {result['score']:.2%}")

# Batch prediction
results = classifier(["Text 1", "Text 2", "Text 3"])

Custom Device Selection

from transformers import pipeline

# Use GPU if available
classifier = pipeline(
    "text-classification",
    model="brightertiger/garbled-text-detector",
    device=0  # Use GPU 0, or -1 for CPU
)

# Use CPU explicitly
classifier = pipeline(
    "text-classification",
    model="brightertiger/garbled-text-detector",
    device=-1
)

Confidence Thresholding

from transformers import pipeline

classifier = pipeline("text-classification", model="brightertiger/garbled-text-detector")

def is_text_valid(text, confidence_threshold=0.85):
    result = classifier(text)[0]
    
    # Accept if classified as NORMAL with high confidence
    if result['label'] == 'NORMAL' and result['score'] > confidence_threshold:
        return True
    
    # Reject if classified as GARBLED with high confidence
    if result['label'] == 'GARBLED' and result['score'] > confidence_threshold:
        return False
    
    # Uncertain - manual review recommended
    return None  # Flag for manual review

Limitations

• Optimized for English text
• May struggle with intentionally stylized text (e.g., l33t speak used artistically)
• Designed for PDF parsing validation; may not generalize to all corruption types
• Max sequence length: 512 tokens

License

Apache 2.0

Citation

@misc{garbled_text_detector_2025,
  title={Garbled Text Detector: BERT-based Binary Classifier for Text Quality Assessment},
  author={BrighterTiger},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/brightertiger/garbled-text-detector}
}

Acknowledgments

• Hugging Face for model hosting and Transformers library
• Huawei Noah's Ark Lab for TinyBERT base model
• PyTorch Lightning for training framework

---

Model Card: brightertiger/garbled-text-detector
Version: 1.0.0
Last Updated: September 2025
Task: Binary Text Classification

Support & Issues

For questions, bug reports, or feature requests:

• 🐛 Open an issue on GitHub
• 💬 Discuss on the Hugging Face Community
• 📧 Contact: [[email protected]]

Contributing

Contributions are welcome! If you'd like to improve the model or add features, please open a pull request on GitHub.